autonomio / talos

Hyperparameter Experiments with TensorFlow and Keras
https://autonom.io
MIT License
1.62k stars 268 forks source link

Future development: TF 2.0 and tf.keras #284

Closed beeb closed 5 years ago

beeb commented 5 years ago

First, thanks to the author, I love Talos and use it every day for my work. I see that TF 2.0 is starting to emerge as alpha and we can certainly expect a release in the next few months. Is there a plan to adapt Talos to the new tf.keras in TF2 in the future?

Thanks!

mikkokotila commented 5 years ago

100%. We must do it. Do you already have a handle of the difference between Keras and tf.keras or is it the same?

EDIT: also thanks a lot for the positive feedback :) I'm glad you benefit from Talos.

mikkokotila commented 5 years ago

This seems to answer my question: https://www.tensorflow.org/guide/keras

This is really an amazing progression I must say. I think TensorFlow has roughly 5x more users than Keras, so many more people can benefit from Talos :)

beeb commented 5 years ago

Amazing news! I don't know if tf.keras differs a lot from normal keras for your purposes. There's a lot of differences under the hood but I think the interface to train the model etc is similar, so my guess is that a lot of your codebase should be compatible. It will require some testing for sure but I hope it's an easy port!

mikkokotila commented 5 years ago

To the extent I can see from examples, it might be really just to test that it works rather than having to change / fix something.

mikkokotila commented 5 years ago

I've looked into this, and it does seem that from v.6.0 onwards, tf.keras is supported exactly like keras is currently. I think, because of the way Talos works, that one should be able to run TensorFlow codes together with a tf.keras model inside input_model for Scan(). In facf, one should be able to run any arbitrary codes as long as the model otherwise follows the Talos way (e.g. keras style history object).

Closing here. Feel free to open a new issue if anything.